Preparation method of tungsten oxide-based memristor and target tracking system

Abstract

The present application relates to a tungsten oxide-based memristor preparation method and a target tracking system, wherein the tungsten oxide-based memristor preparation method comprises the following steps: cleaning indium tin oxide glass for a set time to obtain a bottom electrode; growing a tungsten target on the bottom electrode by direct current magnetron sputtering to obtain a tungsten oxide film and form a first resistance change layer; growing an indium gallium zinc oxide target on the first resistance change layer by radio frequency magnetron sputtering to obtain an indium gallium zinc oxide film and form a second resistance change layer; and obtaining a silver top electrode by covering a mask plate and direct current magnetron sputtering to form a top electrode. The performance of the memristor is stable, the preparation process is simple, the repeatability is high, the thin film uniformity is good, the multi-level resistance state regulation and the modulation of the memristor behavior can be easily realized by changing the preparation parameters and the test parameters, the target tracking system based on the memristor can realize the brain synapse bionic function, and can be widely applied to simulate the human brain to realize higher-level cognitive functions.

Description

Technical Field

[0001] This invention relates to the field of memristor technology, specifically to a method for preparing a tungsten oxide-based memristor and a target tracking system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Memristors are circuit devices that represent the relationship between magnetic flux and charge. They can be used to realize the biomimetic function of brain-like synapses, thereby breaking down the boundary between processing units and storage units, and thus realizing the functions of biometric signal recognition, image processing, and target tracking for biomimetic visual capture.

[0004] Currently, memristors have the following problems:

[0005] (1) Most memristors require expensive materials, high technical costs, and complex manufacturing processes.

[0006] (2) Most memristors have unstable performance, uneven thin film preparation, and low repeatability, which cannot meet the requirements of stable synaptic performance and multiple adjustable resistance states needed to build artificial neural networks.

[0007] (3) The long-term enhancement (LTP) and long-term inhibition (LTD) characteristics of the neural synaptic properties of most memristors are difficult to have high linearity.

[0008] (4) Most memristors are limited to recognizing biometric signals such as electroencephalogram (EEG), electrocardiogram (ECG), sound signals and pulse signals, or image recognition such as handwritten digit recognition, face recognition and image processing for realizing brain-like synaptic bionic functions. They cannot simulate higher-level cognitive functions of the human brain. Summary of the Invention

[0009] To address the technical problems mentioned above, this invention provides a method for fabricating a tungsten oxide-based memristor and a target tracking system. This method ensures stable memristor performance, a simple fabrication process, high repeatability, and good thin-film uniformity. By altering the oxygen atmosphere during fabrication, the valence state of metal ions is changed, while simultaneously increasing the number of oxygen ions and oxygen vacancies, facilitating multi-level resistance state control and modulation of memristor behavior. The fabricated device exhibits high linearity in its LTP and LTD characteristics of the neural synapse. Utilizing this characteristic, an artificial neural network is constructed to achieve target tracking. By inputting a video into a pre-trained neural network model and calibrating the target to be tracked in the first frame of the video, the neural network model can track the target in every frame of the video, including fast-moving targets, occluded targets, and highly similar targets.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The first aspect of the present invention provides a method for preparing a tungsten oxide-based memristor, comprising the following steps:

[0012] The indium tin oxide glass is cleaned for a set time to obtain the bottom electrode;

[0013] A tungsten oxide film is grown on the bottom electrode by DC magnetron sputtering on a tungsten target to form the first resistive switching layer.

[0014] An indium gallium zinc oxide (IGZO) thin film is grown on the first resistive switching layer by radio frequency magnetron sputtering, thus forming the second resistive switching layer.

[0015] The silver target is coated with a template and DC magnetron sputtering to obtain the silver top electrode, forming the top electrode.

[0016] It also includes testing, specifically: testing the prepared tungsten oxide-based memristor under positive and negative voltage pulse conditions based on set parameters.

[0017] It also includes performance tuning, specifically: by changing the fabrication parameters of the resistive switching layer fabrication process and the test voltage pulse parameters of the testing process, the linearity of the long-term duration enhancement (LTP) and long-term duration inhibition (LTD) characteristics in the neural synaptic properties of the tungsten oxide-based memristor is adjusted.

[0018] The bottom electrode is obtained by cleaning the indium tin oxide glass. Specifically, the indium tin oxide glass is ultrasonically cleaned with ethanol and deionized water for a set time to obtain the bottom electrode.

[0019] The fabrication parameters for the resistive switching layer include:

[0020] Ar when obtaining tungsten oxide thin film 2 With O 2 The gas flow rates are 20-30 sccm and 10-20 sccm, respectively.

[0021] The DC power of the tungsten oxide thin film was set to 80-100W.

[0022] The deposition time for the tungsten oxide film was set to 10-20 min.

[0023] Ar to obtain indium gallium zinc oxide thin films 2 With O 2 The gas flow rates are 20-30 sccm and 2-5 sccm, respectively.

[0024] The radio frequency power of the indium gallium zinc oxide thin film was obtained and set to 100-120W;

[0025] The deposition time for obtaining the indium gallium zinc oxide thin film was set to 10-20 min.

[0026] The test voltage pulse parameters during the test process include:

[0027] The positive voltage pulse amplitude is 1.5V-2.3V, and the negative voltage pulse amplitude is -1.3V-2.3V;

[0028] The number of positive and negative voltage pulses is 100-500.

[0029] A second aspect of the present invention provides a tungsten oxide-based memristor prepared by the above method, comprising:

[0030] The bottom electrode, the resistive switching layer, and the top electrode are arranged sequentially from bottom to top. The resistive switching layer includes a first resistive switching layer near the bottom electrode and a second resistive switching layer near the top electrode.

[0031] The bottom electrode is indium tin oxide glass, the first resistive switching layer is a tungsten oxide thin film, the second resistive switching layer is an indium gallium zinc oxide thin film, and the top electrode is a silver top electrode.

[0032] A third aspect of the present invention provides a target tracking system based on the above-described tungsten oxide-based memristor, comprising:

[0033] The backbone network unit is configured to: acquire the current frame and historical frames of the image and extract features, obtain the current frame features and historical frame features and the corresponding foreground label mapping, and input them into the temporally salient attention network unit;

[0034] The temporal salient attention network unit is configured to: acquire the current frame features and historical frame features of the image extracted by the backbone network unit, increase the target feature weights through the target salient module, filter similar features in the current frame and historical frames through the memory module, and concatenate the outputs of the target salient module and the memory module and send them to the classification and regression network unit.

[0035] The classification and regression network unit is configured to: obtain the output of the target saliency module and the memory module concatenated in the temporally salient attention network unit, and determine the target location.

[0036] The backbone network unit, the temporal salient attention network unit, and the classification and regression network unit form a neural network model. The model is trained using a dataset to obtain the required weights. The weights are normalized using the long-term enhancement and long-term inhibition properties of the tungsten oxide-based memristor synapse characteristics and mapped to the weight values ​​in the neural network model for weight replacement. The model is then tested to obtain the trained neural network model.

[0037] The trained neural network model calibrates the target based on the first frame of the video and tracks the target in each frame of the video.

[0038] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects:

[0039] 1. In terms of fabrication process, different types of magnetron sputtering processes are used for the two resistive switching layers and the top electrode, which makes the prepared film uniform, simple, low cost and can be prepared on a large area. This results in the prepared tungsten oxide-based memristor film having high uniformity and good neural synaptic properties.

[0040] 2. In terms of performance, by changing the resistive switching layer structure, the preparation parameters of the resistive switching layer preparation process, and the test voltage pulse parameters of the testing process, the linearity of the long-term duration enhancement (LTP) and long-term duration inhibition (LTD) characteristics in the neural synaptic properties of the tungsten oxide-based memristor was improved, resulting in a target tracking system based on the tungsten oxide-based memristor having good accuracy. Attached Figure Description

[0041] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0042] Figure 1 This is a schematic diagram of a three-dimensional structure of a memristor provided in one or more embodiments of the present invention;

[0043] Figure 2 This is a memristor forward triangular pulse scan IV diagram provided in one or more embodiments of the present invention;

[0044] Figure 3 This is a memristor forward triangular pulse scan Vt diagram provided in one or more embodiments of the present invention;

[0045] Figure 4 This is a memristor negative triangular pulse scan IV diagram provided in one or more embodiments of the present invention;

[0046] Figure 5 This is a memristor negative triangular pulse scan Vt diagram provided in one or more embodiments of the present invention;

[0047] Figure 6 This is a schematic diagram of the pulse amplitude-dependent plasticity of a memristor provided in one or more embodiments of the present invention;

[0048] Figure 7 This is a schematic diagram of the memristor pulse interval dependent plasticity provided by one or more embodiments of the present invention;

[0049] Figure 8 This is a schematic diagram of the pulse width-dependent plasticity of a memristor provided in one or more embodiments of the present invention;

[0050] Figure 9 This is a graph showing the LTP and LTD characteristics of the neural synaptic properties of a memristor provided in one or more embodiments of the present invention.

[0051] Figure 10 This is an application flowchart of a memristor-based target tracking system provided by one or more embodiments of the present invention;

[0052] Figure 11 This is a schematic diagram of the network structure design for a memristor-based target tracking system provided in one or more embodiments of the present invention;

[0053] Figure 12 This is a schematic diagram illustrating the application of a memristor-based target tracking system provided by one or more embodiments of the present invention;

[0054] Figures 13(a)-(c) are schematic diagrams illustrating the video categories that the target tracking system design provided by one or more embodiments of the present invention can recognize;

[0055] In the figure: 1-Tungsten oxide thin film; 2-IGZO thin film; 3-ITO glass; 4-Silver top electrode. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0057] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0058] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0059] As described in the background section, most memristors require expensive materials, involve high technical costs, and have complex fabrication processes. Furthermore, their performance is unstable, the thin film fabrication is uneven, and reproducibility is low, failing to meet the stable synaptic performance and multiple tunable resistive states required for constructing artificial neural networks. The long-term potentiation (LTP) and long-term inhibition (LTD) characteristics of neural synapses do not exhibit high linearity. Moreover, the ability to achieve brain-like synaptic biomimetic functions is limited to recognizing biometric signals such as EEG, ECG, sound, and pulse signals, or image recognition such as handwritten digit recognition, facial recognition, and image processing; it cannot simulate higher-level cognitive functions of the human brain.

[0060] Therefore, the following embodiments provide a method for fabricating tungsten oxide-based memristors and a target tracking system, which ensures stable performance of the memristor, simple fabrication process, high repeatability, and good film uniformity. By changing the oxygen atmosphere during fabrication, the valence state of metal ions is altered, while a large number of oxygen ions and oxygen vacancies are increased, facilitating multi-level resistance state control and modulation of memristor behavior. The LTP and LTD characteristics of the fabricated device's neural synaptic properties exhibit high linearity. Utilizing this characteristic, an artificial neural network is constructed to achieve target tracking. A video is input into a pre-trained neural network model, and the target to be tracked is calibrated in the first frame of the video. The neural network model can then track the target in each frame of the video, and can track fast-moving targets, occluded targets, and highly similar targets.

[0061] Example 1:

[0062] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0063] (1) Commercial indium tin oxide (ITO) glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0064] (2) Preparation of the first resistive switching layer: A tungsten oxide (WO3) film was grown on ITO using a tungsten target with 99.99% purity via DC magnetron sputtering. For this deposition process, the argon (Ar2) / oxygen (O2) gas flow ratio was 30 sccm:15 sccm, the DC power was 80 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0065] (3) Preparation of the second resistive switching layer: An IGZO film was grown on ITO / WO3 using an indium gallium zinc oxide (IGZO) target via radio frequency magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 30 sccm: 2 sccm, the radio frequency power was 100 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0066] (4) Preparation of the top electrode: An Ag top electrode was deposited on the prepared ITO / WO3 / IGZO sample by DC magnetron sputtering using a 99.99% pure silver (Ag) target with a mask over a template. For this deposition process, the Ar2 gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0067] In terms of fabrication process, the magnetron sputtering process has the advantages of producing uniform thin films, simple process, and large-area fabrication, resulting in tungsten oxide-based memristor films with high uniformity and good neural synaptic properties. Furthermore, by modifying the resistive switching layer structure, the fabrication parameters of the resistive switching layer, and the test voltage pulse parameters during testing, the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor was improved.

[0068] Example 2:

[0069] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0070] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0071] (2) Preparation of the first resistive switching layer: A tungsten oxide film was grown on ITO using a tungsten target with 99.99% purity via DC magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 20 sccm:10 sccm, the DC power was 80 W, the deposition pressure was 1.1 Pa, and the deposition time was 10 min.

[0072] (3) Preparation of the second resistive switching layer: An IGZO film was grown on ITO / WO3 using an IGZO target via radio frequency magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 20 sccm: 5 sccm, the radio frequency power was 100 W, the deposition pressure was 1.1 Pa, and the deposition time was 10 min.

[0073] (4) Fabrication of the top electrode: An Ag top electrode was deposited on the prepared ITO / WO3 / IGZO sample using DC magnetron sputtering with a 99.99% pure silver target through a mask-covering method. The resulting memristor structure is shown below. Figure 1 As shown. For this deposition process, the Ar2 gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0074] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0075] Example 3:

[0076] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0077] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0078] (2) Preparation of the first resistive switching layer: A tungsten oxide film was grown on ITO using a tungsten target with 99.99% purity via DC magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 20 sccm:10 sccm, the DC power was 100 W, the deposition pressure was 1.1 Pa, and the deposition time was 10 min.

[0079] (3) Preparation of the second resistive switching layer: An IGZO film was grown on ITO / WO3 using an IGZO target via radio frequency magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 20 sccm: 5 sccm, the radio frequency power was 120 W, the deposition pressure was 1.1 Pa, and the deposition time was 10 min.

[0080] (4) Fabrication of the top electrode: An Ag top electrode was deposited on the prepared ITO / WO3 / IGZO sample using DC magnetron sputtering with a 99.99% pure silver target through a mask-covering method. The resulting memristor structure is shown below. Figure 1 As shown. For this deposition process, the Ar2 gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0081] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0082] Example 4:

[0083] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0084] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0085] (2) Preparation of the first resistive switching layer: A tungsten oxide film is grown on ITO using a 99.99% pure tungsten target via DC magnetron sputtering. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 20 sccm:10 sccm, the DC power was 80 W, the deposition pressure was 1.1 Pa, and the deposition time was 20 min.

[0086] (3) Preparation of the second resistive switching layer: An IGZO film was grown on ITO / WO3 using an IGZO target via radio frequency magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 20 sccm: 5 sccm, the radio frequency power was 100 W, the deposition pressure was 1.1 Pa, and the deposition time was 20 min.

[0087] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1 As shown. For this deposition process, Ar 2 The gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0088] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0089] Example 5:

[0090] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0091] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0092] (2) Preparation of the first resistive switching layer: A tungsten oxide film is grown on ITO using a 99.99% pure tungsten target via DC magnetron sputtering. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 20 sccm:10 sccm, the DC power was 100 W, the deposition pressure was 1.1 Pa, and the deposition time was 20 min.

[0093] (3) Fabrication of the second resistive switching layer: using an IGZO target for radio frequency magnetron sputtering on ITO / WO3 3 An IGZO film is grown on the substrate. For this deposition process, Ar... 2 / O2 The gas flow rate ratio was 20 sccm: 5 sccm, the radio frequency power was 120 W, the deposition pressure was 1.1 Pa, and the deposition time was 20 min.

[0094] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1 As shown. For this deposition process, the Ar2 gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0095] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0096] Example 6:

[0097] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0098] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0099] (2) Preparation of the first resistive switching layer: A tungsten oxide film is grown on ITO using a 99.99% pure tungsten target via DC magnetron sputtering. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm:15 sccm, the DC power was 80 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0100] (3) Fabrication of the second resistive switching layer: using an IGZO target for radio frequency magnetron sputtering on ITO / WO3 3 An IGZO film is grown on the substrate. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm: 2 sccm, the radio frequency power was 100 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0101] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1As shown. For this deposition process, the Ar2 gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0102] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0103] Example 7:

[0104] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0105] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0106] (2) Preparation of the first resistive switching layer: A tungsten oxide film is grown on ITO using a 99.99% pure tungsten target via DC magnetron sputtering. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm:15 sccm, the DC power was 100 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0107] (3) Fabrication of the second resistive switching layer: using an IGZO target for radio frequency magnetron sputtering on ITO / WO3 3 An IGZO film is grown on the substrate. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm: 2 sccm, the radio frequency power was 120 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0108] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1 As shown. For this deposition process, Ar 2 The gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0109] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0110] Example 8:

[0111] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0112] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0113] (2) Preparation of the first resistive switching layer: A tungsten oxide film is grown on ITO using a 99.99% pure tungsten target via DC magnetron sputtering. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm:15 sccm, the DC power was 80 W, the deposition pressure was 1.3 Pa, and the deposition time was 20 min.

[0114] (3) Fabrication of the second resistive switching layer: using an IGZO target for radio frequency magnetron sputtering on ITO / WO3 3 An IGZO film is grown on the substrate. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm: 2 sccm, the radio frequency power was 100 W, the deposition pressure was 1.3 Pa, and the deposition time was 20 min.

[0115] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1 As shown. For this deposition process, Ar 2 The gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0116] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0117] Example 9:

[0118] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0119] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0120] (2) Preparation of the first resistive switching layer: A tungsten oxide film was grown on ITO using a tungsten target with 99.99% purity via DC magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 30 sccm:15 sccm, the DC power was 100 W, the deposition pressure was 1.3 Pa, and the deposition time was 20 min.

[0121] (3) Fabrication of the second resistive switching layer: using an IGZO target for radio frequency magnetron sputtering on ITO / WO3 3 An IGZO film is grown on the substrate. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm: 2 sccm, the radio frequency power was 120 W, the deposition pressure was 1.3 Pa, and the deposition time was 20 min.

[0122] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1 As shown. For this deposition process, Ar 2 The gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0123] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 500 positive voltage pulses with an amplitude of 2V and 500 negative voltage pulses with an amplitude of -1.4V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0124] Example 10:

[0125] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0126] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0127] (2) Preparation of the first resistive switching layer: A tungsten oxide film is grown on ITO using a 99.99% pure tungsten target via DC magnetron sputtering. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm:15 sccm, the DC power was 80 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0128] (3) Preparation of the second resistive switching layer: An IGZO film was grown on ITO / WO3 using an IGZO target via radio frequency magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 30 sccm: 2 sccm, the radio frequency power was 100 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0129] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1 As shown. For this deposition process, Ar 2 The gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0130] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 300 positive voltage pulses with an amplitude of 2.3V and 300 negative voltage pulses with an amplitude of -1.5V, so as to adjust the linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0131] Example 11:

[0132] The method for preparing a tungsten oxide-based memristor includes the following steps:

[0133] (1) Commercial ITO glass was ultrasonically cleaned for 10 min by passing it through ethanol and deionized water in sequence;

[0134] (2) Preparation of the first resistive switching layer: A tungsten oxide film is grown on ITO using a 99.99% pure tungsten target via DC magnetron sputtering. For this deposition process, Ar... 2 / O 2 The gas flow rate ratio was 30 sccm:15 sccm, the DC power was 80 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0135] (3) Preparation of the second resistive switching layer: An IGZO film was grown on ITO / WO3 using an IGZO target via radio frequency magnetron sputtering. For this deposition process, the Ar2 / O2 gas flow rate ratio was 30 sccm: 2 sccm, the radio frequency power was 100 W, the deposition pressure was 1.3 Pa, and the deposition time was 10 min.

[0136] (4) Fabrication of the top electrode: A 99.99% pure silver target was used to perform DC magnetron sputtering on the prepared ITO / WO4 substrate by covering a mask template. 3 The resulting memristor structure is shown below, with Ag top electrode deposited on the IGZO sample. Figure 1As shown. For this deposition process, the Ar2 gas flow rate was 20 sccm, the DC power was 50 W, the deposition pressure was 1 Pa, and the deposition time was 15 min.

[0137] (5) The prepared tungsten oxide-based memristor was tested under the conditions of 100 positive voltage pulses with an amplitude of 1.8V and 100 negative voltage pulses with an amplitude of -1.3V, so as to adjust the linearity of LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor.

[0138] Example 12:

[0139] like Figure 1 As shown, the tungsten oxide-based memristor prepared using the method in Example 1 includes:

[0140] The bottom electrode, the resistive switching layer, and the top electrode are arranged sequentially from bottom to top. The resistive switching layer includes a first resistive switching layer near the bottom electrode and a second resistive switching layer near the top electrode.

[0141] In this embodiment, the bottom electrode is ITO (indium tin oxide) glass 3;

[0142] In this embodiment, the first resistive switching layer is a tungsten oxide thin film 1, and the second resistive switching layer is an IGZO (indium gallium zinc oxide) thin film 2.

[0143] In this embodiment, the top electrode is a silver top electrode 4.

[0144] In this embodiment, the positive triangular pulse scan IV curve of the prepared tungsten oxide-based memristor is as follows: Figure 2 As shown, the positive triangular pulse scan Vt curve is as follows: Figure 3 As shown, the negative triangular pulse scan IV curve is as follows: Figure 4 As shown, the negative triangular pulse scan Vt curve is as follows: Figure 5 As shown; the amplitude of the impulse depends on plasticity, such as Figure 6 As shown, pulse interval depends on plasticity as Figure 7 As shown, pulse width depends on plasticity as Figure 8 As shown.

[0145] In terms of performance, the prepared tungsten oxide-based memristor has the characteristics of high film uniformity and good neural synaptic properties, and the linearity of LTP and LTD is greatly improved.

[0146] The method for adjusting the linearity of the LTP (long-term potentiation) and LTD (long-term inhibition) characteristics of the neural synaptic properties of the aforementioned tungsten oxide-based memristor is as follows:

[0147] (1) Change the resistive switching layer structure of the device;

[0148] (2) Change the preparation parameters of the resistive switching layer preparation process;

[0149] (3) Change the test voltage pulse parameters during the test process.

[0150] In step (1), the structure of the resistive switching layer of the device is changed as follows: after a tungsten oxide thin film is prepared by DC magnetron sputtering, an IGZO thin film is prepared by radio frequency magnetron sputtering.

[0151] In step (2), the preparation parameters of the resistive switching layer preparation process are changed as follows:

[0152] (a) Change the flow rates of Ar2 and O2 gases used in the preparation of tungsten oxide thin films to 20-30 scc m and 10-20 sccm, respectively;

[0153] (b) Change the DC power for preparing tungsten oxide thin films to 80-100W;

[0154] (c) Change the deposition time for preparing tungsten oxide thin films to 10-20 min;

[0155] (d) Change the flow rates of Ar2 and O2 gases during the preparation of IGZO thin films to 20-30 sccm and 2-5 sccm, respectively;

[0156] (e) Change the radio frequency power for preparing the IGZO thin film to 100-120W;

[0157] (f) Change the deposition time for preparing IGZO thin films to 10-20 min.

[0158] In step (3), the test voltage pulse parameters during the test process are changed as follows:

[0159] (a) Change the amplitude of the positive and negative voltage pulses during the test process, setting the positive voltage pulse amplitude to 1.5V-2.3V and the negative voltage pulse amplitude to -1.3V--2.3V;

[0160] (b) Change the number of positive and negative voltage pulses during the test process, setting it to 100-500.

[0161] Example 13:

[0162] The target tracking system based on the tungsten oxide-based memristor of Example Twelve includes:

[0163] The backbone network unit is configured to: acquire the current frame and historical frames of the image and extract features, obtain the current frame features and historical frame features and the corresponding foreground label mapping, and input them into the temporally salient attention network unit;

[0164] The temporal salient attention network unit is configured to: acquire the current frame features and historical frame features of the image extracted by the backbone network unit, increase the target feature weights through the target salient module, filter similar features in the current frame and historical frames through the memory module, and concatenate the outputs of the target salient module and the memory module and send them to the classification and regression network unit.

[0165] The classification and regression network unit is configured to: obtain the output of the target saliency module and the memory module concatenated in the temporally salient attention network unit, and determine the target location.

[0166] In this embodiment:

[0167] First, a neural network model is constructed, consisting of four parts: input, backbone network, temporal salient attention network (TESANet), and classification and regression network.

[0168] The backbone network uses a Siamese neural network framework to extract features from tracking frames and historical frames, and stores the dynamically sampled historical frame features in the memory network. TESANet performs feature interaction on the features extracted by the backbone network, processes the current frame features and historical frame features respectively, and generates features that are easier for subsequent networks to use. The classification and regression network is used to label the detection boxes.

[0169] Secondly, using one NVIDIA GeForce 1080Ti graphics card, the PyTorch deep learning framework was configured in an Ubuntu 16.04 environment to train and test the network model, and the Got-10k dataset was selected for training.

[0170] Finally, by utilizing the characteristics of the LTP and LTD devices, the weight values ​​are normalized and mapped to the weight values ​​in the constructed neural network model to replace the weight values. Then, the model is input into the Got-10k dataset for testing.

[0171] Specifically. For example... Figure 10 As shown:

[0172] (1) A tungsten oxide-based memristor was fabricated. The linearity of the LTP and LTD characteristics in the neural synaptic properties of the tungsten oxide-based memristor was adjusted by changing the fabrication parameters of the resistive switching layer and the test voltage pulse parameters of the testing process. Figure 9 As shown;

[0173] (2) Figure 11 As shown, the entire network structure consists of four parts: input, backbone network, TESANet, and classification and regression network.

[0174] The three historical frames and the current frame are input into the backbone network to extract features from the historical frames and the current frame.

[0175] The backbone network uses the Inception V3 network, but the network parameters are not shared between the two. After historical frame feature extraction, the output consists of three historical frame features and three corresponding foreground and background label mappings L, which are stored in the memory network.

[0176] The current frame undergoes feature extraction through the backbone network and adaptive convolutional layers to adjust the feature dimensions, ultimately serving as input to TESANet along with the memory network.

[0177] After obtaining the features of the current frame and historical frames, TESANet increases the target feature weights through the Target Satisfactory Module (OSM), making the features of the tracked target stand out in the current frame, enhancing the network's ability to learn the local features of the tracked target, and continuously enhancing this effect by cascading OSMs.

[0178] Meanwhile, the memory module (MEM) utilizes multiple sets of historical frame features and, through an information exchange "bridge," filters similar features between the current frame and historical frames. It uses multiple similar features as additional templates to provide extra appearance information, enabling the network to better locate targets in the current frame.

[0179] Finally, the outputs of OSM and MEM are concatenated and fed into subsequent networks for bounding box labeling. An anchorless network is used for bounding box labeling and regression. The entire anchorless network has three branches: classification branch, centrality branch, and regression branch. Each branch first uses a lightweight network with seven convolutions to process the information input from TESANet, and then uses a single convolution to reduce its dimensionality for subsequent classification and regression tasks.

[0180] The classification branch is used for foreground / background classification, determining whether a feature belongs to the target or the background. The centrality branch introduces the FCOS centrality formula to enhance the regression effect of the predicted bounding box. The regression branch directly estimates the target location and ultimately performs the labeling of the detection box.

[0181] (3) Input the Got-10k dataset into the designed neural network model for training to obtain the weights of each part. Utilize the characteristics of the device's LTP and LTD to normalize the weight values ​​and map them to the weight values ​​in the constructed neural network model for weight replacement. Figure 11 As shown, we then input the Got-10k dataset for testing.

[0182] (4) Input the video into the pre-trained neural network model, and label the target to be tracked in the first frame of the video. The neural network model can then track the target in each frame of the video, such as... Figure 12 As shown, it can track fast-moving targets, occluded targets, and highly similar targets, such as... Figures 13(a)-13(c) As shown.

[0183] The target tracking system based on tungsten oxide memristors has good accuracy. After adjusting the linearity of the LTP and LTD characteristics in the neural synaptic properties of tungsten oxide memristors, the output signal changes more uniformly with the input signal. When performing weight replacement, it can make the selected quantized conductance state and the weights of the constructed artificial neural network model form a more ideal mapping relationship, thereby improving the accuracy.

[0184] In this embodiment, there are three evaluation metrics for accuracy: the average overlap rate between the ground truth and the detection box (AO), the detection success rate with an overlap threshold of 0.5 (SR0.5), and the detection success rate with an overlap threshold of 0.75 (SR0.75). Taking AO as an example, the calculation process for the evaluation metrics is as follows:

[0185]

[0186] Where C represents different video categories in the dataset, SC represents the images in C, and |SC| is the number of images in C. As shown in formula (1), firstly, the average overlap rate (AO) of all images in a certain video category is summed and the average value is calculated to obtain the mAO value of a single video category. Then, the mAO values ​​of all video categories are summed and averaged to obtain the final mAO. The calculation methods of SR0.5 and SR0.75 are the same as mAO. Compared with SR0.5, SR0.75 has stricter requirements on the tracker. On the Got-10k official website, mAO is the primary evaluation metric, and mAO is equivalent to AO.

[0187] In this embodiment, the AO is 65.6, the SR0.5 is 75.7, and the SR0.75 is 59, which are very close to the accuracy generated using a computer model.

[0188] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for preparing a tungsten oxide-based memristor, characterized in that, Includes the following steps: The indium tin oxide glass is cleaned for a set time to obtain the bottom electrode; A tungsten oxide film is grown on the bottom electrode by DC magnetron sputtering on a tungsten target to form the first resistive switching layer. An indium gallium zinc oxide (IGZO) thin film is grown on the first resistive switching layer by radio frequency magnetron sputtering, thus forming the second resistive switching layer. The silver target is coated with a template and DC magnetron sputtering to obtain the silver top electrode, forming the top electrode.

2. The method for preparing a tungsten oxide-based memristor as described in claim 1, characterized in that, It also includes testing, specifically: The prepared tungsten oxide-based memristor was tested under positive and negative voltage pulse conditions based on set parameters.

3. The method for preparing a tungsten oxide-based memristor as described in claim 1, characterized in that, This also includes performance tuning, specifically: By changing the fabrication parameters of the resistive switching layer and the test voltage pulse parameters of the testing process, the linearity of the long-term enhancement and long-term inhibition characteristics in the neural synaptic properties of the tungsten oxide-based memristor can be adjusted.

4. The method for preparing a tungsten oxide-based memristor as described in claim 1, characterized in that, The bottom electrode is obtained by cleaning the indium tin oxide glass, specifically as follows: The bottom electrode is obtained by ultrasonically cleaning the indium tin oxide glass with ethanol and deionized water for a set time.

5. The method for preparing a tungsten oxide-based memristor as described in claim 1, characterized in that, The parameters for fabricating resistive switching layers include: The flow rates of Ar2 and O2 gases when the tungsten oxide thin film was obtained were 20-30 sccm and 10-20 sccm, respectively. The DC power of the tungsten oxide thin film was set to 80-100W. The deposition time for the tungsten oxide film was set to 10-20 min. Ar to obtain indium gallium zinc oxide thin films 2 With O 2 The gas flow rates are 20-30 sccm and 2-5 sccm, respectively. The radio frequency power of the indium gallium zinc oxide thin film was obtained and set to 100-120W; The deposition time for obtaining the indium gallium zinc oxide thin film was set to 10-20 min.

6. The method for preparing a tungsten oxide-based memristor as described in claim 2, characterized in that, The test voltage pulse parameters during the test process include: The positive voltage pulse amplitude is 1.5V-2.3V, and the negative voltage pulse amplitude is -1.3V-2.3V; The number of positive and negative voltage pulses is 100-500.

7. A tungsten oxide-based memristor obtained by the preparation method according to any one of claims 1-6, characterized in that, include: The bottom electrode, the resistive switching layer, and the top electrode are arranged sequentially from bottom to top. The resistive switching layer includes a first resistive switching layer near the bottom electrode and a second resistive switching layer near the top electrode. The bottom electrode is indium tin oxide glass, the first resistive switching layer is a tungsten oxide thin film, the second resistive switching layer is an indium gallium zinc oxide thin film, and the top electrode is a silver top electrode.

8. A target tracking system based on the tungsten oxide-based memristor as described in claim 7, characterized in that, include: The backbone network unit is configured to: acquire the current frame and historical frames of the image and extract features, obtain the current frame features and historical frame features and the corresponding foreground label mapping, and input them into the temporally salient attention network unit; The temporal salient attention network unit is configured to: acquire the current frame features and historical frame features of the image extracted by the backbone network unit, increase the target feature weights through the target salient module, filter similar features in the current frame and historical frames through the memory module, and concatenate the outputs of the target salient module and the memory module and send them to the classification and regression network unit. The classification and regression network unit is configured to: obtain the output of the target saliency module and the memory module concatenated in the temporally salient attention network unit, and determine the target location.

9. The target tracking system as described in claim 8, characterized in that, The backbone network unit, the temporal salient attention network unit, and the classification and regression network unit form a neural network model. The model is trained using a dataset to obtain the required weights. The weights are normalized using the long-term enhancement and long-term inhibition properties of the tungsten oxide-based memristor synapse characteristics and mapped to the weight values ​​in the neural network model for weight replacement. The model is then tested to obtain the trained neural network model.

10. The target tracking system as described in claim 9, characterized in that, The trained neural network model calibrates the target based on the first frame of the video and tracks the target in each frame of the video.

Images